A generative model for evaluating missing data methods in large epidemiological cohorts.

Abstract

Background: The potential value of large scale datasets is constrained by the ubiquitous problem of missing data, arising in either a structured or unstructured fashion. When imputation methods are proposed for large scale data, one limitation is the simplicity of existing evaluation methods. Specifically, most evaluations create synthetic data with only a simple, unstructured missing data mechanism which does not resemble the missing data patterns found in real data. For example, in the UK Biobank missing data tends to appear in blocks, because non-participation in one of the sub-studies leads to missingness for all sub-study variables.

Methods: We propose a tool for generating mixed type missing data mimicking key properties of a given real large scale epidemiological data set with both structured and unstructured missingness while accounting for informative missingness. The process involves identifying sub-studies using hierarchical clustering of missingness patterns and modelling the dependence of inter-variable correlation and co-missingness patterns.

Results: On the UK Biobank brain imaging cohort, we identify several large blocks of missing data. We demonstrate the use of our tool for evaluating several imputation methods, showing modest accuracy of imputation overall, with iterative imputation having the best performance. We compare our evaluations based on synthetic data to an exemplar study which includes variable selection on a single real imputed dataset, finding only small differences between the imputation methods though with iterative imputation leading to the most informative selection of variables.

Conclusions: We have created a framework for simulating large scale data with that captures the complexities of the inter-variable dependence as well as structured and unstructured informative missingness. Evaluations using this framework highlight the immense challenge of data imputation in this setting and the need for improved missing data methods.